Lebesgue Outer Measure and the Carathéodory Construction

Intuition Beginner

Imagine you have a strange shape on the real line, perhaps the set of all rational numbers between $0$ and $1$, or perhaps the set of points where some complicated function jumps. You want to assign it a size. The cleanest idea is also the oldest: cover the shape with intervals whose total length you can compute, and let the size be the smallest total length you can squeeze the cover down to. That smallest cover-length is the outer measure of the shape.

Outer measure is an upper-bound rule. It gives every subset of the line a number between zero and infinity. It always assigns the empty set zero. Larger shapes get larger numbers. If you cut a shape into countably many pieces, the outer measure of the whole is at most the sum of the outer measures of the pieces. Those three properties are the full definition, and they hold for every subset, no exceptions.

The surprise is that outer measure is not yet a measure. Suppose you take two disjoint subsets of the line, $A$ and $B$. You might hope the outer measure of $A\cup B$ equals the outer measure of $A$ plus the outer measure of $B$. For many disjoint pairs this holds, but for pathological pairs it can fail. Carathéodory's brilliant move was to pick out, from the universe of all subsets, the well-behaved ones for which additivity does hold. His test: a set $A$ is well-behaved when, for every other subset $T$ at all, $A$ slices $T$ cleanly into two pieces whose outer measures sum back to the outer measure of $T$.

The collection of well-behaved sets forms a sigma-algebra, the same kind of measurable-set structure from the previous unit. Restricted to that sigma-algebra, outer measure stops being a mere upper bound and becomes an actual measure: it is countably additive, not just countably sub-additive. This restricted measure on the line is exactly Lebesgue measure.

The one-sentence takeaway: outer measure is the easy first guess at size, the Carathéodory test is the filter that turns the easy guess into a real measure on the sets that pass.

Visual Beginner

Picture a wiggly subset $A$ of the real line drawn as a horizontal band of dots. Above it, draw several covers: first a coarse cover by three wide intervals, then a refined cover by ten narrower intervals, then a very fine cover by hundreds of tiny intervals. The total length of each cover shrinks as the cover tightens. The infimum of these total lengths is the outer measure of $A$.

The bottom panel shows the Carathéodory test in action. A test set $T$ is cut by a candidate $A$ into $T\cap A$ and the leftover. If the outer measures of those two pieces always add back to the outer measure of $T$, then $A$ passes the test and earns a place in the Lebesgue sigma-algebra.

Worked example Beginner

We compute the outer measure of the rational numbers in $[0,1]$, which we will call $A$.

Step 1. List the rationals in $A$ as a sequence $q_1,q_2,q_3,\dots$. This is possible because the rationals are countable.

Step 2. Pick a small target $\epsilon >0$. Around the rational $q_n$, place an open interval of length $\epsilon /2^n$, so the interval covers $q_n$ and stretches half that length to each side.

Step 3. The union of these intervals covers every rational in $A$, since each $q_n$ sits inside its own interval. The total length of the cover is the sum $\epsilon /2+\epsilon /4+\epsilon /8+\cdots$, which equals $\epsilon$.

Step 4. The outer measure of $A$ is at most $\epsilon$. Since $\epsilon$ was an arbitrary positive number, the outer measure of $A$ is $0$.

What this tells us: the rationals, despite being dense in $[0,1]$, have outer measure zero. Density and size do not measure the same thing.

Check your understanding Beginner

Formal definition Intermediate+

Let $X$ be a non-empty set and let $\mathcal{P}(X)$ denote its power set.

Definition (outer measure). An outer measure on $X$ is a function ${\mu }^{\ast }:\mathcal{P}(X)\to [0,\infty ]$ satisfying ^{[Folland §1.4]}:

1. ${\mu }^{\ast }(\varnothing )=0$.
2. (Monotonicity.) If $A\subseteq B$, then ${\mu }^{\ast }(A)\le {\mu }^{\ast }(B)$.
3. (Countable sub-additivity.) If $\{A_n{\}}_{n\in \mathbb{N}}$ is a countable family of subsets of $X$, then ${\mu }^{\ast }\left({\bigcup }_{n=1}^{\infty }{A}_n\right)\le {\sum }_{n=1}^{\infty }{\mu }^{\ast }(A_n)$.

The defining feature of an outer measure, as distinct from a measure, is that the domain is the entire power set, with no requirement of additivity for disjoint unions.

Definition (Lebesgue outer measure on ${\mathbb{R}}^n$). For any $A\subseteq {\mathbb{R}}^n$, set $$m^{\ast }(A)=\inf \left\{\sum _{k=1}^{\infty }\mathrm{vol}(R_k):A\subseteq \bigcup _{k=1}^{\infty }{R}_k,R_k\text{ open box}\right\},$$ where an open box is a product ${\prod }_{i=1}^n(a_i,b_i)$ and $\mathrm{vol}(R_k)={\prod }_{i=1}^n(b_i-a_i)$.

When $n=1$, the open boxes are open intervals and $\mathrm{vol}$ reduces to length. The infimum is taken over all countable open-box covers; finite covers appear as the special case of a countable cover with all but finitely many terms empty.

Definition (Carathéodory measurability). Let ${\mu }^{\ast }$ be an outer measure on $X$. A subset $A\subseteq X$ is Carathéodory-$\mu^$-measurable\ast (orsimply\ast$\mu^$-measurable) when $${\mu }^{\ast }(T)={\mu }^{\ast }(T\cap A)+{\mu }^{\ast }(T\setminus A)\text{for every }T\subseteq X.$$

The collection of ${\mu }^{\ast }$-measurable sets is denoted $\mathcal{M}({\mu }^{\ast })$ ^{[Carathéodory 1914]}.

Since $T=(T\cap A)\cup (T\setminus A)$ is a disjoint union, sub-additivity gives ${\mu }^{\ast }(T)\le {\mu }^{\ast }(T\cap A)+{\mu }^{\ast }(T\setminus A)$ automatically. The substantive direction is the reverse inequality, so the measurability test reduces to $${\mu }^{\ast }(T)\ge {\mu }^{\ast }(T\cap A)+{\mu }^{\ast }(T\setminus A)\text{for every }T\subseteq X.$$

The Lebesgue sigma-algebra on ${\mathbb{R}}^n$ is $\mathcal{L}=\mathcal{M}(m^{\ast })$, and Lebesgue measure is $m=m^{\ast }{|}_{\mathcal{L}}$. The triple $({\mathbb{R}}^n,\mathcal{L},m)$ is the canonical Lebesgue measure space.

Counterexamples to common slips Intermediate+

• Outer measure is not a measure. The Vitali set $V\subseteq [0,1]$ from 02.07.01 is not $m^{\ast }$-measurable; the additivity test fails for a specific $T$. Outer measure remains defined on $V$ — it is just not countably additive there.
• The covering family matters. The infimum is over countable covers, not finite covers. Restricting to finite open-cover infima gives Jordan outer content, which differs from Lebesgue outer measure on sets such as the rationals in $[0,1]$.
• Closed and open boxes give the same value. Replacing open boxes with closed boxes, or with half-open boxes, in the cover family yields the same outer measure. The choice is a matter of convention; the resulting infimum is unchanged because boundary faces have outer measure zero.

Key theorem with proof Intermediate+

Theorem (Carathéodory's extension theorem). Let ${\mu }^{\ast }$ be an outer measure on a set $X$. Then $\mathcal{M}({\mu }^{\ast })$ is a sigma-algebra, and the restriction $\mu ={\mu }^{\ast }{|}_{\mathcal{M}({\mu }^{\ast })}$ is a complete measure on $\mathcal{M}({\mu }^{\ast })$ ^{[Carathéodory 1914]}.

Proof. We verify the three sigma-algebra axioms for $\mathcal{M}({\mu }^{\ast })$ and countable additivity for $\mu$.

Axiom 1. $X\in \mathcal{M}({\mu }^{\ast })$. For any test set $T$, $T\cap X=T$ and $T\setminus X=\varnothing$, so ${\mu }^{\ast }(T\cap X)+{\mu }^{\ast }(T\setminus X)={\mu }^{\ast }(T)+0={\mu }^{\ast }(T)$.

Axiom 2. $\mathcal{M}({\mu }^{\ast })$ is closed under complementation. The defining identity is symmetric in $A$ and $A^c=X\setminus A$, since $T\cap A^c=T\setminus A$ and $T\setminus A^c=T\cap A$.

Closure under finite unions. Let $A,B\in \mathcal{M}({\mu }^{\ast })$ and let $T$ be any test set. Apply the measurability of $A$ to $T$: $${\mu }^{\ast }(T)={\mu }^{\ast }(T\cap A)+{\mu }^{\ast }(T\setminus A).$$ Apply the measurability of $B$ to $T\setminus A$: $${\mu }^{\ast }(T\setminus A)={\mu }^{\ast }((T\setminus A)\cap B)+{\mu }^{\ast }((T\setminus A)\setminus B).$$ Combine and rearrange using $T\cap (A\cup B)=(T\cap A)\cup ((T\setminus A)\cap B)$ and $T\setminus (A\cup B)=(T\setminus A)\setminus B$: $${\mu }^{\ast }(T)={\mu }^{\ast }(T\cap A)+{\mu }^{\ast }((T\setminus A)\cap B)+{\mu }^{\ast }(T\setminus (A\cup B)).$$ Sub-additivity applied to $T\cap (A\cup B)=(T\cap A)\cup ((T\setminus A)\cap B)$ gives ${\mu }^{\ast }(T\cap (A\cup B))\le {\mu }^{\ast }(T\cap A)+{\mu }^{\ast }((T\setminus A)\cap B)$. Hence $${\mu }^{\ast }(T)\ge {\mu }^{\ast }(T\cap (A\cup B))+{\mu }^{\ast }(T\setminus (A\cup B)),$$ and the reverse inequality is automatic, so $A\cup B\in \mathcal{M}({\mu }^{\ast })$.

Finite additivity. If $A,B\in \mathcal{M}({\mu }^{\ast })$ are disjoint, take $T=A\cup B$ in the measurability of $A$: $${\mu }^{\ast }(A\cup B)={\mu }^{\ast }((A\cup B)\cap A)+{\mu }^{\ast }((A\cup B)\setminus A)={\mu }^{\ast }(A)+{\mu }^{\ast }(B).$$ By induction, ${\mu }^{\ast }({⨆}_{k=1}^n{A}_k)={\sum }_{k=1}^n{\mu }^{\ast }(A_k)$ for any finite disjoint family in $\mathcal{M}({\mu }^{\ast })$.

Axiom 3. Closure under countable disjoint unions. Let $\{A_n{\}}_{n\ge 1}$ be a disjoint family in $\mathcal{M}({\mu }^{\ast })$ and set $B_n={⨆}_{k=1}^n{A}_k$ and $B={⨆}_{k=1}^{\infty }{A}_k$. Each $B_n\in \mathcal{M}({\mu }^{\ast })$ by finite-union closure. For any test set $T$, $${\mu }^{\ast }(T)={\mu }^{\ast }(T\cap B_n)+{\mu }^{\ast }(T\setminus B_n)\ge {\mu }^{\ast }(T\cap B_n)+{\mu }^{\ast }(T\setminus B),$$ since $T\setminus B\subseteq T\setminus B_n$ and ${\mu }^{\ast }$ is monotone. Iterating finite additivity inside $T\cap B_n$: $${\mu }^{\ast }(T\cap B_n)=\sum _{k=1}^n{\mu }^{\ast }(T\cap A_k).$$ So $${\mu }^{\ast }(T)\ge \sum _{k=1}^n{\mu }^{\ast }(T\cap A_k)+{\mu }^{\ast }(T\setminus B)\text{for every }n.$$ Letting $n\to \infty$ and applying sub-additivity in reverse: $${\mu }^{\ast }(T)\ge \sum _{k=1}^{\infty }{\mu }^{\ast }(T\cap A_k)+{\mu }^{\ast }(T\setminus B)\ge {\mu }^{\ast }(T\cap B)+{\mu }^{\ast }(T\setminus B),$$ where the second inequality uses sub-additivity for $T\cap B={⨆}_k(T\cap A_k)$. The reverse direction is automatic. Hence $B\in \mathcal{M}({\mu }^{\ast })$. Closure under arbitrary countable unions follows because any countable union can be rewritten as a countable disjoint union via $A_n^{\prime }=A_n\setminus {\bigcup }_{k<n}{A}_k$, and complements/finite intersections preserve $\mathcal{M}({\mu }^{\ast })$.

Countable additivity of $\mu$. Setting $T=B$ in the inequality chain above yields ${\mu }^{\ast }(B)\ge {\sum }_k{\mu }^{\ast }(A_k)$, and sub-additivity gives the reverse, so $\mu ({⨆}_k{A}_k)={\sum }_k\mu (A_k)$ for disjoint families in $\mathcal{M}({\mu }^{\ast })$.

Completeness. If ${\mu }^{\ast }(N)=0$ and $A\subseteq N$, then for any test set $T$, sub-additivity gives ${\mu }^{\ast }(T\cap A)\le {\mu }^{\ast }(A)\le {\mu }^{\ast }(N)=0$, so ${\mu }^{\ast }(T\cap A)=0$, and ${\mu }^{\ast }(T\setminus A)\le {\mu }^{\ast }(T)$ by monotonicity. The reverse ${\mu }^{\ast }(T)\le {\mu }^{\ast }(T\cap A)+{\mu }^{\ast }(T\setminus A)$ from sub-additivity completes the equality. Hence every ${\mu }^{\ast }$-null subset is in $\mathcal{M}({\mu }^{\ast })$, and $\mu$ is a complete measure.

$\square$

Bridge. The Carathéodory construction builds toward 02.07.04 the Lebesgue integral, where measurability of the underlying sets transfers to measurability of functions $f:(X,\mathcal{M}({\mu }^{\ast }))\to (\mathbb{R},\mathcal{B}(\mathbb{R}))$ and unlocks the monotone and dominated convergence theorems. The foundational reason the construction works is that the measurability test propagates through complementation and countable disjoint union by a sub-additivity argument applied to slices, and this is exactly the structure that identifies the Lebesgue sigma-algebra with the completion of the Borel sigma-algebra under $m^{\ast }$-null sets. The bridge is between two pictures of the same object: the descriptive Borel hierarchy from 02.07.01 generated by open intervals, and the analytic Carathéodory sigma-algebra extracted from the outer-measure test. Putting these together, the central insight is that completion under null sets corresponds to passage from generators to the full ${\mu }^{\ast }$-measurable class; this pattern generalises to every measure-theoretic extension theorem.

Exercises Intermediate+

Lean formalization Intermediate+

lean_status: full — Mathlib provides the full Carathéodory construction. MeasureTheory.OuterMeasure encodes outer measures on a type. MeasureTheory.OuterMeasure.caratheodory constructs the sigma-algebra of ${\mu }^{\ast }$-measurable sets. MeasureTheory.OuterMeasure.ofFunction builds an outer measure from an arbitrary set function via the cover-and-take-infimum recipe used in this unit. MeasureTheory.volume is Lebesgue measure on real Euclidean space, defined by exactly this Carathéodory recipe applied to the box-volume pre-measure. The companion module records the unit's predicate names and confirms the construction is in scope.

import Mathlib.MeasureTheory.OuterMeasure.Defs
import Mathlib.MeasureTheory.OuterMeasure.Caratheodory
import Mathlib.MeasureTheory.OuterMeasure.OfFunction
import Mathlib.MeasureTheory.Measure.Lebesgue.Basic

variable (X : Type*)

abbrev CodexOuterMeasure : Type _ := MeasureTheory.OuterMeasure X

def CodexCaratheodory (mu : MeasureTheory.OuterMeasure X) :
    MeasurableSpace X :=
  mu.caratheodory

abbrev CodexLebesgue (n : Nat) : MeasureTheory.Measure (Fin n -> Real) :=
  MeasureTheory.volume

Advanced results Master

The advanced theory of the Carathéodory construction splits into three strands: the abstract extension theorem (from pre-measures on semi-rings to measures on sigma-algebras), the structure of Lebesgue measure on ${\mathbb{R}}^n$ (completeness, regularity, translation-invariant uniqueness), and the foundational role of the axiom of choice (Vitali, Banach-Tarski, Solovay).

Theorem 1 (Hahn-Kolmogorov extension theorem). Let $\mathcal{A}$ be an algebra of subsets of $X$ and let ${\mu }_0:\mathcal{A}\to [0,\infty ]$ be a pre-measure — that is, ${\mu }_0(\varnothing )=0$, ${\mu }_0$ is countably additive on countable disjoint families whose union is in $\mathcal{A}$, and ${\mu }_0$ is $\sigma$-finite (there is a countable $\{X_n\}\subseteq \mathcal{A}$ with $X=\bigcup X_n$ and ${\mu }_0(X_n)<\infty$). Then ${\mu }_0$ extends uniquely to a measure $\mu$ on $\sigma (\mathcal{A})$. The extension is achieved by the outer measure ${\mu }^{\ast }(A)=\inf \{{\sum }_n{\mu }_0(A_n):A\subseteq \bigcup A_n,A_n\in \mathcal{A}\}$, followed by the Carathéodory restriction to $\mathcal{M}({\mu }^{\ast })\supseteq \sigma (\mathcal{A})$ ^{[Kolmogorov 1933]}.

The existence half is a direct application of Carathéodory's extension theorem from the Intermediate proof. Uniqueness uses the monotone class theorem (or $\pi$-$\lambda$) inside each $X_n$ and assembles. This is the foundational existence + uniqueness result for measure theory: every classical measure (Lebesgue, Lebesgue-Stieltjes, Hausdorff, Haar, product measures) is constructed by specifying a pre-measure on an algebra and invoking Hahn-Kolmogorov.

Theorem 2 (Lebesgue measure is the unique translation-invariant Borel measure with unit cube one). On $({\mathbb{R}}^n,\mathcal{B}({\mathbb{R}}^n))$, Lebesgue measure $m$ is characterised by: (i) translation-invariance $m(E+x)=m(E)$; (ii) $m([0,1{]}^n)=1$; (iii) finiteness on compact sets. Any Borel measure with these properties equals $m$ ^{[Folland §1.5]}.

The proof is Exercise 8 above. Translation-invariance plus a single normalisation pins the measure down on dyadic cubes, hence on rational-corner boxes, hence on all open sets by countable union, hence on all Borel sets by the $\pi$-$\lambda$ theorem.

Theorem 3 (Cardinality of $\mathcal{L}$ versus $\mathcal{B}({\mathbb{R}}^n)$). $|\mathcal{B}({\mathbb{R}}^n)|=2^{{\aleph }_0}$ and $|\mathcal{L}|=2^{2^{{\aleph }_0}}$.

The cardinality of the Borel sigma-algebra follows from transfinite induction on the Borel hierarchy, as in 02.07.01. The Lebesgue cardinality jumps because the Cantor set $C\subseteq [0,1]$ is a Borel null set of cardinality $2^{{\aleph }_0}$, and by completeness every subset of $C$ is Lebesgue-measurable (with measure zero). So $\mathcal{P}(C)\subseteq \mathcal{L}$ contributes $2^{2^{{\aleph }_0}}$ sets, and the immediate upper bound $|\mathcal{L}|\le |\mathcal{P}({\mathbb{R}}^n)|=2^{2^{{\aleph }_0}}$ closes the count. The Saks 1937 theorem ^{[Saks 1937]} expresses this as: $\mathcal{L}$ modulo the ideal of null sets has the same cardinality as $\mathcal{P}(\mathbb{R})$.

Theorem 4 (Regularity of Lebesgue measure). For every Lebesgue-measurable set $E\subseteq {\mathbb{R}}^n$, $$m(E)=\inf \{m(U):U\supseteq E,U\text{ open}\}=\sup \{m(K):K\subseteq E,K\text{ compact}\}.$$ Equivalently, $\mathcal{L}$ is the smallest sigma-algebra containing the Borel sets and all subsets of Borel null sets, and every measurable set $E$ can be written as $E=B△N$ for some Borel $B$ and some $N$ contained in a Borel null set ^{[Folland §1.5]}.

Outer regularity comes from the definition of $m^{\ast }$ via open-box covers. Inner regularity uses outer regularity on the complement within a bounded box: $m(E)=m([-N,N{]}^n)-m([-N,N{]}^n\setminus E)$ for bounded $E$, with the complement's outer-regular approximation by opens corresponding to inner-regular approximation by compact subsets of $E$.

Theorem 5 (Hausdorff measures). For each $s\ge 0$, the $s$-dimensional Hausdorff outer measure on ${\mathbb{R}}^n$ is $${\mathcal{H}}^s(A)=\underset{\delta \to 0^{+}}{\lim }{\mathcal{H}}_{\delta }^s(A),{\mathcal{H}}_{\delta }^s(A)=\inf \left\{\sum _k(\mathrm{diam}{U}_k{)}^s:A\subseteq \bigcup U_k,\mathrm{diam}{U}_k<\delta \right\}.$$ For each $A$ there is a unique critical exponent ${\dim }_H(A)\in [0,n]$ — the Hausdorff dimension of $A$ — such that ${\mathcal{H}}^s(A)=\infty$ for $s<{\dim }_H(A)$ and ${\mathcal{H}}^s(A)=0$ for $s>{\dim }_H(A)$. The Hausdorff measure is constructed by Carathéodory's method and is countably additive on its associated Carathéodory sigma-algebra, which contains every Borel set ^{[Federer 1969 §2.10]}.

The Cantor set has Hausdorff dimension $\log 2/\log 3\approx 0.6309$, computed by self-similarity: $C$ is the disjoint union of two scaled copies of itself at ratio $1/3$, and the equation $2\cdot (1/3{)}^s=1$ yields $s=\log 2/\log 3$. Hausdorff measures underpin geometric measure theory and fractal geometry ^{[Federer 1969 §2.10]}.

Theorem 6 (Vitali non-measurable set under choice). Assuming the axiom of choice, there exists $V\subseteq [0,1]$ with $V\notin \mathcal{L}$ ^{[Vitali 1905]}.

The construction is Exercise 7. The Vitali set highlights that the Lebesgue sigma-algebra, while large, is not all of $\mathcal{P}(\mathbb{R})$, and the gap is irreducible under full choice.

Theorem 7 (Solovay's consistency theorem). Assuming the consistency of ZFC plus an inaccessible cardinal, the theory ZF + DC (dependent choice) + "every subset of $\mathbb{R}$ is Lebesgue-measurable" is consistent ^{[Solovay 1970]}.

In Solovay's model, the Vitali construction fails because the axiom of choice is replaced by dependent choice, which is too weak to pick representatives from uncountable equivalence relations. The model also satisfies "every subset of $\mathbb{R}$ has the Baire property" and "every uncountable subset of $\mathbb{R}$ contains a perfect set." These three classical regularity properties — measurability, Baire property, perfect-set property — are jointly consistent with ZF + DC. The existence of non-measurable sets is therefore a strictly weaker assumption than full AC, but not implied by ZF + DC.

Theorem 8 (Banach-Tarski paradox). In ${\mathbb{R}}^3$, assuming the axiom of choice, there exist a partition of the unit ball $B\subseteq {\mathbb{R}}^3$ into finitely many pieces (five suffice) and a sequence of rotations and translations that reassemble the pieces into two disjoint copies of $B$ ^{[Banach-Tarski 1924]}.

The proof uses the free group $F_2$ on two generators acting on $\mathrm{SO}(3)$, a paradoxical decomposition of $F_2$, the axiom of choice to choose orbit representatives, and the failure of Lebesgue-measurability of the pieces. The pieces are necessarily non-Lebesgue-measurable; this is the geometric content of non-measurability in three dimensions. Banach-Tarski does not contradict Lebesgue measure because the measure is undefined on the pieces; it dramatises the cost of demanding both rigid-motion invariance and additivity for every subset.

The two-dimensional case behaves differently. In ${\mathbb{R}}^2$, the group of orientation-preserving rigid motions is solvable, and by a theorem of Banach (1923) there exists a finitely-additive, rotation-invariant extension of Lebesgue measure to every subset of the plane. The paradox is therefore a feature of dimension $n\ge 3$, where the rigid-motion group is non-amenable in the von Neumann sense and contains free subgroups. The free-subgroup structure of $\mathrm{SO}(3)$ — explicitly, two rotations about distinct axes by suitable angles generate a free group, as shown by Hausdorff 1914 in his sphere-paradox paper — is the algebraic engine of the construction.

Synthesis. The Carathéodory construction is the foundational reason that the measure-theoretic framework scales beyond the real line and beyond Lebesgue measure. The central insight is that an outer measure plus the slice-additivity test extracts a sigma-algebra and a measure simultaneously, and this is exactly the structure that identifies pre-measures on algebras with measures on the generated sigma-algebras through Hahn-Kolmogorov. Putting these together with the translation-invariant uniqueness theorem and the regularity theorem, Lebesgue measure on ${\mathbb{R}}^n$ emerges as a canonical object: pinned down by symmetry and a single normalisation, automatically regular from both sides, automatically complete.

The pattern generalises in three directions at once. First, replacing box-volume with $(\mathrm{diam}{)}^s$ in the cover infimum produces Hausdorff measures, which probe sets of fractional dimension and provide the bridge to geometric measure theory and fractal analysis. Second, replacing the algebra of half-open boxes with an algebra of cylinder sets in a product space produces product measures and the Fubini-Tonelli theorem; this construction appears again in probability theory 02.26.01 pending and in the construction of Haar measure on locally compact groups. Third, allowing the pre-measure to come from a distribution function $F:\mathbb{R}\to \mathbb{R}$ produces Lebesgue-Stieltjes measures, which encode probability distributions and underpin stochastic-process construction (Kolmogorov 1933). The bridge is between the analytic Carathéodory criterion and the probabilistic axiomatisation, and Saks 1937 closed the loop by recognising both as instances of a single extension theorem.

The role of the axiom of choice in producing non-measurable sets (Vitali) and in producing the Banach-Tarski paradox identifies Lebesgue measure as the largest object on which we can demand both translation-invariance and countable additivity under standard set theory. Solovay's 1970 consistency result shows that the existence of non-measurable sets is itself a marker of how much choice we are willing to use; dropping AC to DC erases the Vitali pathology, but at the cost of losing other classical theorems whose pattern recurs throughout twentieth-century analysis.

Full proof set Master

Proposition 1 ($m^{\ast }$ is an outer measure on ${\mathbb{R}}^n$). The Lebesgue outer measure $m^{\ast }$ defined by the box-cover infimum satisfies $m^{\ast }(\varnothing )=0$, monotonicity, and countable sub-additivity.

Proof. For $m^{\ast }(\varnothing )=0$: the family with a single empty cover gives infimum zero.

For monotonicity: if $A\subseteq B$, every cover of $B$ is a cover of $A$, so $m^{\ast }(A)\le m^{\ast }(B)$.

For countable sub-additivity: let $\{A_k{\}}_{k\ge 1}$ be subsets and fix $\epsilon >0$. For each $k$, choose a box cover $\{R_{k,j}{\}}_{j\ge 1}$ of $A_k$ with ${\sum }_j\mathrm{vol}(R_{k,j})\le m^{\ast }(A_k)+\epsilon /2^k$. The doubly-indexed family $\{R_{k,j}{\}}_{k,j}$ covers ${\bigcup }_k{A}_k$ and is countable (countable union of countable families). Hence $$m^{\ast }\left(\bigcup _k{A}_k\right)\le \sum _{k,j}\mathrm{vol}(R_{k,j})=\sum _k\sum _j\mathrm{vol}(R_{k,j})\le \sum _k{m}^{\ast }(A_k)+\epsilon .$$ Letting $\epsilon \to 0$ gives the inequality.

$\square$

Proposition 2 (Lebesgue measure of a box). For any half-open box $R={\prod }_{i=1}^n[a_i,b_i)$ with $a_i<b_i$, $m^{\ast }(R)={\prod }_{i=1}^n(b_i-a_i)=\mathrm{vol}(R)$.

Proof. Upper bound: the single open box $\prod (a_i-\epsilon ,b_i+\epsilon )$ covers $R$ and has volume $\prod (b_i-a_i+2\epsilon )\to \mathrm{vol}(R)$ as $\epsilon \to 0$.

Lower bound: this is the one-dimensional Heine-Borel argument from Exercise 4 raised to dimension $n$. Let $\{R_k\}$ be any open-box cover of $R$. Shrink $R$ slightly to a compact sub-box $R^{\prime }=\prod [a_i+\delta ,b_i-\delta ]$ for small $\delta >0$. By compactness, finitely many $R_{k_1},\dots ,R_{k_N}$ already cover $R^{\prime }$. The sum ${\sum }_{j=1}^N\mathrm{vol}(R_{k_j})$ is bounded below by $\mathrm{vol}(R^{\prime })$ by an inclusion-exclusion + projection-volume argument (alternatively, by induction on $n$). Hence ${\sum }_k\mathrm{vol}(R_k)\ge \mathrm{vol}(R^{\prime })$, and as $\delta \to 0$ this gives ${\sum }_k\mathrm{vol}(R_k)\ge \mathrm{vol}(R)$. Infimum over covers gives $m^{\ast }(R)\ge \mathrm{vol}(R)$.

$\square$

Proposition 3 (Carathéodory completeness). $({\mathbb{R}}^n,\mathcal{L},m)$ is complete: if $N\in \mathcal{L}$ with $m(N)=0$ and $A\subseteq N$, then $A\in \mathcal{L}$ and $m(A)=0$.

Proof. By monotonicity, $m^{\ast }(A)\le m^{\ast }(N)=0$, so $m^{\ast }(A)=0$. For any test set $T$, $m^{\ast }(T\cap A)\le m^{\ast }(A)=0$ and $m^{\ast }(T\setminus A)\le m^{\ast }(T)$. Sub-additivity gives $m^{\ast }(T)\le m^{\ast }(T\cap A)+m^{\ast }(T\setminus A)=m^{\ast }(T\setminus A)$. Combining, $$m^{\ast }(T)=m^{\ast }(T\cap A)+m^{\ast }(T\setminus A),$$ so $A\in \mathcal{M}(m^{\ast })=\mathcal{L}$ with $m(A)=m^{\ast }(A)=0$.

$\square$

Proposition 4 (translation invariance of $m^{\ast }$ on ${\mathbb{R}}^n$). For every $A\subseteq {\mathbb{R}}^n$ and every $x\in {\mathbb{R}}^n$, $m^{\ast }(A+x)=m^{\ast }(A)$, and $\mathcal{L}$ is closed under translation.

Proof. Open-box volume is translation-invariant: translating a box $\prod (a_i,b_i)$ by $x$ produces the box $\prod (a_i+x_i,b_i+x_i)$ of the same volume. Hence any open-box cover of $A$ translates to an open-box cover of $A+x$ with the same total volume, and infimum gives $m^{\ast }(A+x)\le m^{\ast }(A)$. The reverse inequality follows by translating by $-x$.

For sigma-algebra closure: if $E\in \mathcal{L}$, then for any test set $T$, $m^{\ast }(T)=m^{\ast }(T\cap E)+m^{\ast }(T\setminus E)$. Replacing $T$ by $T+x$ and using translation-invariance on both outer measures (applied separately to $T\cap E$, $T\setminus E$, $T$ — each translated by $x$): $$m^{\ast }(T)=m^{\ast }(T-x)=m^{\ast }((T-x)\cap E)+m^{\ast }((T-x)\setminus E)=m^{\ast }(T\cap (E+x))+m^{\ast }(T\setminus (E+x)),$$ so $E+x\in \mathcal{L}$.

$\square$

Proposition 5 (Lebesgue sigma-algebra is the completion of Borel under $m$-null sets). $\mathcal{L}=\{B△N:B\in \mathcal{B}({\mathbb{R}}^n),N\subseteq B^{\prime }\text{ for some Borel }{B}^{\prime }\text{ with }m(B^{\prime })=0\}$.

Proof. The inclusion $\supseteq$: every Borel set is in $\mathcal{L}$ (Exercise 6), every subset of a Borel null set is in $\mathcal{L}$ by completeness (Proposition 3), and $\mathcal{L}$ is closed under symmetric difference.

The inclusion $\subseteq$: let $E\in \mathcal{L}$. By outer regularity (Theorem 4 in Advanced results), there exist open sets $U_n\supseteq E$ with $m(U_n)\to m^{\ast }(E)$, so $G={\bigcap }_n{U}_n$ is a Borel set ($G_{\delta }$) containing $E$ with $m(G)=m^{\ast }(E)$. By inner regularity applied to bounded $E$, there exist compact sets $K_n\subseteq E$ with $m(K_n)\to m^{\ast }(E)$ (for unbounded $E$, intersect with cubes $[-N,N{]}^n$ and take a union); $F={\bigcup }_n{K}_n$ is a Borel set ($F_{\sigma }$) contained in $E$ with $m(F)=m^{\ast }(E)$. So $F\subseteq E\subseteq G$, both $F$ and $G$ are Borel, $m(G\setminus F)=0$, and $E△F\subseteq G\setminus F$ is a subset of a Borel null set. Hence $E=F△(E△F)$ is in the right-hand side.

$\square$

Connections Master

• σ-algebra and Borel σ-algebra 02.07.01. The previous unit establishes the sigma-algebra framework, the Borel sigma-algebra generated by open sets, and the existence of non-Borel sets via cardinality. The current unit takes that framework as input and produces Lebesgue measure as the canonical translation-invariant measure on the Borel sigma-algebra plus the completion under null sets. The Lebesgue sigma-algebra $\mathcal{L}$ is exactly the Carathéodory sigma-algebra of $m^{\ast }$, and Proposition 4 identifies it with the Borel completion. The Vitali construction from 02.07.01 reappears here at full strength to demonstrate that $\mathcal{L}\ne \mathcal{P}({\mathbb{R}}^n)$.

• Lebesgue integral and convergence theorems 02.07.04. The downstream unit defines the Lebesgue integral of measurable functions via simple functions and monotone limits, then proves the monotone convergence theorem, Fatou's lemma, and the dominated convergence theorem. Each of these theorems requires the underlying measure space to be complete and countably additive, properties guaranteed by Carathéodory's construction here. The Carathéodory completeness is exactly what licences the "almost everywhere" language: a function modified on a null set is integrable iff the original is, with the same integral.

• $L^p$ spaces and completeness 02.07.06. Building toward 02.07.06, the $L^p$ Banach spaces are defined as equivalence classes of measurable functions modulo equality almost everywhere with respect to Lebesgue measure. The construction depends on the completeness of $({\mathbb{R}}^n,\mathcal{L},m)$: the equivalence relation modding out null sets requires Carathéodory completeness for the resulting quotient to be a Banach space. The Riesz-Fischer theorem identifying $L^2({\mathbb{R}}^n)$ as a Hilbert space sits on this measure-theoretic foundation.

• Banach spaces 02.11.04. The shipped Banach space unit develops completeness, boundedness, and continuous linear functionals. The $L^p$ spaces from 02.07.06 are the canonical analytic examples; their completeness (the Riesz-Fischer theorem) pulls in the Lebesgue-measure foundation built here, since the proof passes through pointwise almost-everywhere convergence on a subsequence, which only makes sense in a complete measure space. The dual-space theory for $L^p$ (Riesz representation: $(L^p{)}^{\ast }\cong L^{p^{\prime }}$ for $1\le p<\infty$) similarly depends on the regularity and translation-invariance theorems proved here.

• Real number axioms 02.02.01. The Heine-Borel compactness argument used in Proposition 2 to compute $m^{\ast }([a,b])=b-a$ relies on the least-upper-bound property of $\mathbb{R}$, which is the analytic content of the 02.02.01 axiomatisation. Without least-upper-bound, the finite-subcover step fails, and the lower bound on box outer-measure collapses. The Lebesgue measure construction is therefore not just topology-driven but order-completeness-driven; the same structure that produces convergent monotone sequences produces convergent decreasing covers.

Historical & philosophical context Master

Lebesgue's 1902 thesis ^{[Lebesgue 1902]} introduced outer measure on the real line as the cover-based infimum we use today, alongside an inner-measure dual constructed from closed subsets. Lebesgue defined a bounded set $E\subseteq \mathbb{R}$ to be measurable when its outer and inner measures agreed, and showed that this collection forms a sigma-algebra containing every Borel set. The Lebesgue integral was then built on this measure space, and the monotone and dominated convergence theorems followed. The thesis was a breakthrough partly because it resolved the convergence problem that had plagued Riemann integration: a pointwise limit of Riemann-integrable functions need not be Riemann-integrable, but the analogous statement for Lebesgue-integrable functions holds under modest hypotheses.

The conceptual lineage running into Lebesgue's thesis traces back to Borel's 1898 Leçons sur la théorie des fonctions, which assigned a measure to the Borel sigma-algebra via countable additivity over disjoint open intervals, and further back to Peano (1887) and Jordan (1892), whose Peano-Jordan content captured the finitely-additive theory but failed for dense countable sets such as the rationals. Lebesgue's innovation was to replace finite covers with countable covers and replace the Peano-Jordan inner-outer-content gap with a finer Lebesgue inner-outer-measure gap that closed for far more sets.

Carathéodory's 1914 paper ^{[Carathéodory 1914]} reframed Lebesgue's outer-and-inner approach with a single slice-additivity criterion: $A$ is measurable iff every test set $T$ is split cleanly by $A$ in the outer-measure sense. The Carathéodory criterion eliminated the need for an independently defined inner measure and worked in arbitrary set-function settings, not just for the box-volume pre-measure on ${\mathbb{R}}^n$. Carathéodory's 1918 book ^{[Carathéodory 1918]} presented the criterion as the foundation of an abstract theory of measure, distinct from Lebesgue's geometric origins.

Banach and Tarski's 1924 paper ^{[Banach-Tarski 1924]} produced the geometric paradox bearing their names, using the axiom of choice to partition a 3-ball into finitely many pieces that reassemble into two copies of the ball. The paradox dramatised what Vitali had shown in 1905 ^{[Vitali 1905]}: under full AC, finitely-additive rigid-motion-invariant measures cannot extend Lebesgue measure to every subset of ${\mathbb{R}}^n$ for $n\ge 3$. The two-dimensional case is more delicate (a Banach-Tarski paradox in ${\mathbb{R}}^2$ requires both AC and rigid-motion-plus-affine; pure rigid motions in dimension two admit a finitely-additive invariant extension by amenability of the affine group).

Kolmogorov's 1933 monograph ^{[Kolmogorov 1933]} crystallised the measure-theoretic foundations of probability around the Carathéodory extension theorem, identifying probability spaces $(\Omega ,\mathcal{F},P)$ with measure spaces of total mass one. The Hahn-Kolmogorov extension theorem became the workhorse for constructing probability measures on path spaces, leading directly to the modern theory of stochastic processes. Saks's 1937 Theory of the Integral ^{[Saks 1937]} gave the now-standard exposition of outer-measure-based integration and proved the cardinality-of-$\mathcal{L}$-modulo-null result that places the Lebesgue sigma-algebra as the natural completion object.

Solovay's 1970 theorem ^{[Solovay 1970]} settled the foundational question about the axiom of choice: in a model of ZF + DC built from an inaccessible cardinal, every subset of $\mathbb{R}$ is Lebesgue-measurable. The existence of non-measurable sets is therefore not a theorem of ZF + DC but requires choice principles beyond DC. The Solovay model also satisfies the Baire-property and perfect-set analogues, identifying Lebesgue measurability as one of three parallel regularity properties — measurability, Baire property, perfect-set property — that classical analysis tacitly invokes whenever it works on $\mathbb{R}$.

The Solovay model has subsequently been refined by Shelah (1984), who showed that the inaccessible-cardinal hypothesis is necessary for the Lebesgue-measurability conclusion but not for the Baire-property conclusion: the consistency of "every set is Baire" is provable from ZFC alone, while the consistency of "every set is Lebesgue-measurable" requires an inaccessible. This asymmetry is the modern descriptive-set-theoretic content of the Vitali construction and a landmark of the calibration of choice principles.

Lebesgue measure has become the analyst's bedrock. Every $L^p$ space sits over it 02.07.06; every Sobolev space is constructed via weak derivatives integrated against $m$; distribution theory pairs test functions against measures; the spectral theorem for self-adjoint operators rests on spectral measures defined via the Carathéodory extension; the Riesz representation theorem identifies positive linear functionals on continuous functions of compact support with regular Borel measures, and the Lebesgue measure case is the canonical example. Integration-theoretic PDE construction — weak solutions, energy methods, variational formulations — all sit on this Carathéodory-built foundation. The pattern recurs throughout twentieth- and twenty-first-century analysis: state a problem in geometric or analytic language, lift it to a measure-theoretic question, apply the Carathéodory machinery, and recover regularity by sigma-algebra completion.

Bibliography Master

@article{Lebesgue1902,
  author  = {Lebesgue, Henri},
  title   = {Int\'egrale, longueur, aire},
  journal = {Annali di Matematica Pura ed Applicata},
  volume  = {7},
  year    = {1902},
  pages   = {231--359}
}

@article{Caratheodory1914,
  author  = {Carath\'eodory, Constantin},
  title   = {{\"U}ber das lineare Mass von Punktmengen --- eine Verallgemeinerung des L\"angenbegriffs},
  journal = {Nachr. K\"onigl. Ges. Wiss. G\"ottingen, Math.-Phys. Kl.},
  year    = {1914},
  pages   = {404--426}
}

@book{Caratheodory1918,
  author    = {Carath\'eodory, Constantin},
  title     = {Vorlesungen \"uber reelle Funktionen},
  publisher = {Teubner},
  address   = {Leipzig},
  year      = {1918}
}

@article{BanachTarski1924,
  author  = {Banach, Stefan and Tarski, Alfred},
  title   = {Sur la d\'ecomposition des ensembles de points en parties respectivement congruentes},
  journal = {Fundamenta Mathematicae},
  volume  = {6},
  year    = {1924},
  pages   = {244--277}
}

@book{Kolmogorov1933,
  author    = {Kolmogorov, Andrey N.},
  title     = {Grundbegriffe der Wahrscheinlichkeitsrechnung},
  publisher = {Springer},
  year      = {1933}
}

@book{Saks1937,
  author    = {Saks, Stanis{\l}aw},
  title     = {Theory of the Integral},
  edition   = {2},
  publisher = {Hafner},
  year      = {1937}
}

@misc{Vitali1905,
  author    = {Vitali, Giuseppe},
  title     = {Sul problema della misura dei gruppi di punti di una retta},
  publisher = {Tip. Gamberini e Parmeggiani},
  address   = {Bologna},
  year      = {1905}
}

@article{Solovay1970,
  author  = {Solovay, Robert M.},
  title   = {A model of set-theory in which every set of reals is Lebesgue measurable},
  journal = {Annals of Mathematics},
  volume  = {92},
  year    = {1970},
  pages   = {1--56}
}

@book{Halmos1950,
  author    = {Halmos, Paul R.},
  title     = {Measure Theory},
  publisher = {Van Nostrand},
  year      = {1950},
  note      = {Reprinted as Springer GTM 18, 1974}
}

@book{Folland1999,
  author    = {Folland, Gerald B.},
  title     = {Real Analysis: Modern Techniques and Their Applications},
  edition   = {2},
  publisher = {Wiley},
  year      = {1999}
}

@book{Royden2010,
  author    = {Royden, H. L. and Fitzpatrick, P. M.},
  title     = {Real Analysis},
  edition   = {4},
  publisher = {Pearson},
  year      = {2010}
}

@book{Bogachev2007,
  author    = {Bogachev, Vladimir I.},
  title     = {Measure Theory, Volume 1},
  publisher = {Springer},
  year      = {2007}
}

@book{Federer1969,
  author    = {Federer, Herbert},
  title     = {Geometric Measure Theory},
  publisher = {Springer},
  year      = {1969}
}

@book{Tao2016,
  author    = {Tao, Terence},
  title     = {Analysis II},
  edition   = {3},
  publisher = {Hindustan Book Agency},
  year      = {2016}
}